Why Financial Agents Need Domain Capabilities, Not Necessarily a New Framework
A report argues that financial Agent systems should first build domain capabilities such as correct product identities, calculation semantics, evidence chains, and authorization boundaries. Using a teaching portfolio-research scenario, it compares the roles of Pi, Claude Agent SDK, and LangGraph, while distinguishing domain capabilities, execution frameworks, and business products. It recommends starting with mature general-purpose tools and separately evaluating whether a dedicated financial framework or custom execution runtime provides enough value. The portfolio, funds, amounts, and dates in the example are instructional constructs, not actual holdings, and the report is not a cross-framework performance benchmark.
The report provides a practical basis for teams choosing between extending mature general-purpose tools and investing in a dedicated framework or custom runtime, while emphasizing that financial requirements should not be treated as automatic justification for rebuilding execution infrastructure.
` 同步校注（2026-11-04）：文中涉及撤权、来源消费和恢复的历史案例，应与当前 Run 授权生命周期区分。当前实现中，Host 已授予的权限在同一 Run 及其恢复中持续有效，不因外部权限表变化重新撤销；新 Run 独立准入。账户、scope、来源范围、内容身份、预算和显式取消仍按原合同执行。历史撤权复现用于解释当时的问题，不代表当前 Run 内重新授权机制。
原则篇｜Pi、Claude Agent SDK、LangGraph 与金融 Agent 的职责、取舍和退出条件。
当一个 Agent 已经能够搜索网页、阅读报告、执行 Python、调用子代理、保存会话，我们很容易产生一个判断：再做金融垂类框架，是否只是在成熟工具外面换了一层名字？
这个怀疑是合理的。一个系统接上行情接口、加几段金融 Prompt、注册几个基金 Skill，并不因此拥有新的架构价值。反过来，金融业务确实存在专门的事实、计算和责任问题，也不能因为通用框架功能丰富，就把这些问题当成模型会自动处理的细节。